arXiv 论文提出信号碰撞惩罚层：最大回撤从 23.6% 降至 14.1%
Penalize the Common Trade: A Collision-Aware Layer for Learned Portfolios
量化圈的朋友可以看看这篇：多个模型学到同一方向会互相踩踏，论文给了个简单惩罚层，实测把最大回撤砍了近 10 个百分点。
arXiv 论文（2611.03453）研究多个学习系统共享数据或目标后做出近乎相同交易的问题，作者称之为信号碰撞。基于线性影响博弈的分析表明每个配置者会过度承担共同暴露，重叠依赖的收费机制可以恢复社会最优选择。在 10 只流动性跨资产 ETF、2012 年 1 月至 2026 年 7 月共 760 周的冻结面板、10 个基点的自融资成本设定下，在组合构建前减去近期共同收益模式负载的惩罚层，将最大回撤从 23.6% 降至 14.1%，波动率从 12.6% 降至 10.7%，超额收益 Sharpe 从 0.664 变为 0.667。成对平稳自助法的 95% 置信区间较宽，回撤改善区间为 [-3.3, 18.4] 个百分点，包含零。
Penalize the Common Trade: A Collision-Aware Layer for Learned Portfolios
Different learning systems can make nearly the same trade when they share data, objectives, or representations. The market then sees decision concentration even if the models look diverse. We call exposure to this shared direction a signal collision. A linear-impact game shows the basic externality: each allocator chooses too much common exposure because it does not bear the full price effect imposed on others. An overlap-dependent toll recovers the social choice, and a feedback model gives the exact gain at which return-responsive allocation becomes unstable. We also state the restrictive spectral conditions under which a leading eigenportfolio can reveal the collision direction. These results motivate a simple post-model layer: subtract a penalty for loading on the recent common return mode before forming the portfolio. In a frozen panel of ten liquid cross-asset ETFs, 760 completed weeks from January 2012 through July 2026, and exact self-financing costs of 10 basis points per dollar traded, this layer changes maximum drawdown from 23.6% to 14.1%, volatility from 12.6% to 10.7%, and excess-return Sharpe from 0.664 to 0.667. A volatility-matched momentum diagnostic still draws down 19.9%. Paired stationary-bootstrap uncertainty is wide: the 95% interval for drawdown improvement is $[-3.3,18.4]$ percentage points and includes zero. The result establishes a reproducible risk-control association and specifies the holdings-and-flow validation needed to attribute the common direction to AI crowding. AI-use disclosure. Generative AI tools, including OpenAI Codex, assisted manuscript preparation. For this preprint version, Codex assisted with LaTeX editing, author and venue metadata, compilation, and layout checks.